Appearance
Cicero documentation
Start from what you're trying to do. Every guide below is current unless it says otherwise; the historical records at the bottom are provenance, not guidance.
Understand it
| Read | For |
|---|---|
| Project README | What Cicero is, what it feels like, and the quickstart |
| Concepts | The two layers: Cicero itself, and your office in your own config |
| Architecture | How a spoken turn flows, and what runs where |
| Why not full-duplex | The core design decision: honest turn-taking with fast barge-in |
Have your first conversation
| Read | For |
|---|---|
| Setup | The one install path: the setup wizard, step by step, to the first spoken reply |
| Using Cicero | What to say, interrupting, "details", "think hard", privacy day to day, checking what runs |
| Choosing a brain | Which agent to plug in (Claude Code, Codex, Gemini, ACP, any OpenAI-compatible endpoint) and how |
| Configuration | Privacy, escalation, tier presets, the config schema, quick intents, custom voice actions — with config.yaml.example as the annotated reference |
Advanced / example deployment
| Read | For |
|---|---|
| Advanced overview | One owner's office as an example: what it is and what it is not |
| Reference deployment | The owner's always-on box, as a worked example |
| The office | Lanes: several agents with their own voices behind one call |
| Channels | How voice and text reach Cicero, and adding a new channel |
Operate it
| Read | For |
|---|---|
| Web voice | The browser/PWA surface: controls, identity, limits |
| Daemon mode & local mic | Lifecycle, activation, echo cancellation, computer use on the local machine |
| Voice activation | Hands-free start, claps, VAD tuning, earcons |
| Turn detection | Semantic end-of-turn (Smart-Turn): what it fixes and how to enable it |
| The classifier backend | An optional small model held apart from the reply model for per-utterance decisions |
| The intent judge | "Was that addressed to me?" — an opt-in LLM veto over captured speech in a noisy room |
| Voice cloning | Giving Cicero (or a lane) any voice from one reference clip |
| TTS coalescing | Fewer synthesis calls per reply — measured, and off by default |
| Notifications | Cicero speaking up on its own: Telegram, briefings, schedules, quiet hours |
| Telegram calls | The phone-call sidecar: talk to your agent from anywhere |
| Security | Threat model, authentication, egress rules — read before exposing anything beyond localhost |
| What leaves the box | The complete data-flow map: what stays local, what's opt-in, and what Cicero never does |
Extend it
| Read | For |
|---|---|
| Choosing a brain → custom drivers | The three tiers of adding an agent, from zero-code to a small adapter |
| Configuration → quick intents | Your own zero-latency phrases and voice actions, pure YAML |
| Python model servers | How the speech sidecars are provisioned |
History (dated records)
Point-in-time snapshots kept for provenance. Claims inside describe their date, not the current product.
- Lessons learned — the first three days (March 2026)
- Performance & portability evaluation — June 2026 audit
- Evaluation follow-up — the July 2026 hardening series
superpowers/plans/,superpowers/specs/— historical design records, not the backlog